From lab to newsroom: How Reuters builds AI tools journalists actually use

Overview

Reuters developed a suite of AI-powered newsroom tools with a human-in-the-loop approach: Fact Genie, an AI-assisted summarisation tool that scans entire documents in under five seconds to suggest newsworthy alerts (helping Speed teams publish the first alert within about six seconds of a press release, across roughly 100,000 business news alerts published monthly by 250-300 journalists); LEON, an AI-powered headline assistant; and AVISTA, a machine-learning tool for sourcing, tagging and archiving photos and videos. Fact Genie took about four months from prototype to production rollout, starting with senior journalists in the US and UK before expanding, and the team also uses AI to filter out non-newsworthy content before it reaches a large language model, cutting alert-generation time from about a minute with GPT-3.5 Turbo to about 10 seconds with GPT-4o mini.

The challenge

Reuters' Speed teams historically faced the challenge of processing large volumes of information quickly to publish around 100,000 business news alerts monthly across 250-300 journalists. While structured data like earnings could be automated, events like CEO firings or layoffs required ultra-fast human reporting measured in fractions of seconds, with each region having its own reporting style and sources adding to the challenge.

The solution

Reuters developed a suite of newsroom AI tools with a human-in-the-loop approach: Fact Genie, an AI-assisted summarisation tool that scans entire documents in under five seconds to suggest newsworthy alerts; LEON, an AI-powered headline assistant; and AVISTA (Automated Video/Image Sourcing, Tagging and Archiving), which uses machine learning to help journalists quickly find, tag and edit photos and videos. The team also uses AI to filter out non-newsworthy content before it reaches a large language model, cutting alert-generation time from about a minute with GPT-3.5 Turbo to about 10 seconds with GPT-4o mini. Reuters' Bangalore newsroom, now its largest globally, has emerged as a key hub for AI-driven journalism.

Reported business value

Fact Genie scans entire documents in under five seconds to suggest newsworthy alerts, helping Speed teams publish the first alert within about six seconds of a press release. Junior journalists were able to work much faster and meet Reuters' standards more easily, making the work more accessible to them. Filtering non-newsworthy content before sending it to a large language model cut alert-generation time from about a minute with GPT-3.5 Turbo to about 10 seconds with GPT-4o mini.